VTC-KSVD:一种融合视觉特征与标签一致性的多标签图像标注方法  

VTC-KSVD,a New Multi-Label Image Annotation Method Combining Visual Features with Tag Consistency

作　　者：张菊莉[1] 贺占庄[1] 戴涛[2] 张君毅 ZHANG Ju-li;HE Zhan-zhuang;DAI Tao;ZHANG Jun-yi(Xi’an Microelectronics Technique Institute,Xi’an,Shaanxi 710068,China;School of Software Engineering,Xi’an Jiaotong University,Xi’an,Shaanxi 710045,China)

机构地区：[1]西安微电子技术研究所,陕西西安710068 [2]西安交通大学软件学院,陕西西安710045

出　　处：《北京理工大学学报》2020年第2期175-181,188,共8页Transactions of Beijing Institute of Technology

基　　金：中国青年科学基金资助项目(61702413);中国航天九院技术创新基金资助项目(2016JY06)。

摘　　要：提出一种融合视觉特征及标签一致性的多标签图像标注方法VTC-KSVD.首先通过K均值奇异值分解(KSVD)法建立图像的标签一致性模型TC-KSVD,然后将多视图特征融合在该模型中.该方法既利用了训练样本的类标与编码系数的判别式模型,又利用了训练样本的标签与编码系数的关系,增加了字典的判别性,提高了标注性能.在Corel5K数据集上的实验结果表明,融合了多视图视觉特征与标签一致性的VTC-KSVD方法可以较为准确地找到视觉特征与语义特征均相似的图像近邻,能明显提升多标签图像的标注性能,并能有效缓解训练数据有限而引起的稀疏性问题.A new method for multi-label image annotation was proposed based on the combination of visual features and tag consistency.Firstly,a tag consistency model TC-KSVD was established for the training images using the KSVD method.In order to further improve the annotation accuracy,multi-view visual features were incorporated into the model.This method was arranged not only to utilize the discriminant model of the training samples’item labels and coding coefficients,but also to utilize the relationship between tags and the coding coefficients,so as to increase the discriminability of the dictionary and improve the annotation performance.The experimental results on the Corel5K datasets show that,the VTC-KSVD method with multi-view visual features and tag consistency can accurately find the neighbors with similar visual features and semantic features,which can significantly improve the annotation accuracy and can effectively alleviate the sparsity problem caused by limited training data.

关 键 词：图像标注 KSVD 视觉特征 标签一致性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]